On the nature of cavities on protein surfaces: Application to the Identification of drug-binding sites Murad Nayal, Barry Honig Columbia University, NY Proteins: Structure, Function and Bioinformatics, Accepted 15 Nov. 05 Ankur Dhanik

Abstract Identification of drug-binding sites useful for virtual screening and drug design. Small ligands are known to bind proteins at surface cavities. Two tasks: identification of cavities and prediction of their drugabbilities (whether the cavity is suitable for drug binding). The method presented in this paper encoded in program called SCREEN (Surface Cavity REcognition and EvaluatioN).

Abstract SCREEN works by first constructing two molecular surfaces using GRASP: a conventional molecular surface (MS) using a 1.4 A radius and a second low resolution envelope using a large probe sphere, which serves as ‘sea-level’. Depth of each vertex of MS is computed and compared with threshold. For each surface cavity, 408 attributes are computed (physiochemical, structural, and geometric). Random Forests based classifier is used. Training data set is derived from a collection of 100 nonredundant protein ligand complexes.

Results SCREEN predicts drug binding cavities with a balanced error rate of 7.2% and coverage of 88.9%, while a CASTp ( a popular protein cavity detection program) based druggability predictor (using cavity size criteria alone) predicts with a balanced error rate of 15.7% and coverage of 71.7%. SCREEN predicts drug-binding cavities missed by cavity size criteria (three examples). Out 18 attributes out of 408 used, were found to be significant predictors of drug binding cavities. It follows from the above that drug binding cavities are large, deep, have an intricate curvature profile, are rigid, and have a relatively small number of prolines, as well as amino acids with small but negative octanol-to-water transfer free energies (Asn, Gln, Glu).

Results Protein-tyrosine phosphatase 1B, PTP1B (PDB code: 1l8g). The largest surface cavity (colored green: area, 184 Å 2 ; volume, 400 Å 3 ; residues Gln78, Arg79, Ser80, and Pro210) is about 20 Å from the ligand-binding site. The drug binds at the second largest cavity, colored red, as predicted (area, 170 Å 2 ; volume, 259 Å 3 ; residues Gln262, Ala217, Ile219, Val49, and Asp181).

Results Human carbonic anhydrase II (CA II). The largest cavity (area, 281 Å 2 ; volume, 679 Å 3 ; residues Phe213, Tyr7, Gly8, Asp243, and Lys170), shown in green, is rather shallow and is predicted not to bind a drug. Instead, the second largest cavity (area, 194 Å 2 ; volume, 281 Å 3 ; residues Leu198, Thr200, His94, Val121, and His64) is the one predicted correctly to bind the drug.

Results Human factor Xa complexed with inhibitor RPR (PDB code: 1ezq). Four cavities ranked 1, 2, 3, and 9, shown here in red, were predicted to be potential drug-binding cavities. The ligand actually binds at two cavities, 3 (the S1 pocket: area, 274 Å2; volume, 384 Å3; residues Gln192, Trp215, Ser195, Cys191, Gly216, and Asp) and 9 (the S4 pocket: area, 69 Å2; volume, 155 Å3; residues Trp215, Phe174, Thr98, Tyr99, and Ile175).

Results Surface cavity propertyCategory Drug-binding cavities Non drug- binding cavities Cavity rankSize1.89 ± ± 5.4 Number of residuesSize22.8 ± ± 5.4 Number of atomsSize85.0 ± ± 21.2 Smallest moment of inertiaSize/shape1.7 × 10 4 ± 2.5 × × 10 3 ± 8.3 × 10 3 Depth standard deviationSize/shape2.3 ± 1.1 (Å 3 )0.75 ± 0.45 Maximum depthSize/shape10.5 ± 4.0 (Å)4.75 ± 1.67 Average depthSize/shape5.3 ± 1.9 (Å)3.2 ± 0.7 Normalized smallest moment of inertia Shape17.0 ± ± 5.3 Proportion of cavity at depth between [6.5, 6.75) Shape0.02 ± ± Largest moment of inertiaSize/shape1.6 × 10 4 ± 8.4 × × 10 3 ± 1.6 × 10 4 Average side-chains residual entropy Rigidity-0.41 ± 0.18 (kcal) ± 0.25 Average curvatureShape-49.0 ± ± 13.1 Maximum curvednessShape6.4 ± ± 4.9 Maximum mean curvatureShape5.3 ± ± 4.2 Curvedness < 0.5Shape0.35 ± ± 0.08 Proportion of prolineAmino acid composition ± ± 0.09 Proportion of cavity with logP between [-1, 0) Hydrophobicity0.09 ± ± 0.16 Side-chain residual entropy standard deviation Rigidity0.43 ± 0.18 (kcal) 0.55 ± 0.17

Comments The prediction of drug-binding cavities was done without considering the nature of the drug. Physicochemical cavity properties were not found useful. Perhaps they can play an important role when surface cavities that recognize a particular ligand are characterized. Energy-based approach offers a promising alternative to geometry-based methods